Release v4.0.1 #852

Patch release. Two PRs change the binary: #842 (bad configs return errors instead of panicking) and #846 (model loading about 0.6 s faster per run). Output is unchanged at the same seed. Everything else since v4.0.0 is harness, docs or baselines.

Per PR

Binary changes

• Bad configs return errors instead of panicking #842 — bad configs return errors instead of panicking (exit 1, the reason on stderr).
  • Owed: subcommands outside the three config loaders were not audited case by case for user-input panics (they share create_output_file, which is fixed).
• Load models by decompressing first, then parsing (#845) #846 (Model loading: from_reader on the 20.8 MB default error model costs ~0.7 s per run #845) — every model load decompresses before parsing. Startup 0.72 → 0.09 s locally; output byte-identical to develop.
  • Delta re-run: benchmark 22985666 on d06b3a0. E. coli 12.28 → 10.99 s, yeast 24.03 → 22.94 s, chr22 68.20 → 67.44 s. Only chr22's reps are tight enough to quote, because of a possible Delta filesystem incident that weekend (Per-base generation is 11-16% slower than v1.19.1 (same node, local disk); thread scaling is not regressed #848).
  • Owed: the ~40 MB of fixed RSS (Peak RSS carries a fixed ~40 MB overhead that postdates the perf baseline #670) is untouched.

Harness, docs, baselines (no binary change)

• v4.0.0 follow-ups: conda sha256, complete CHANGELOG Not verified list #831 — v4.0.0 conda sha256 and CHANGELOG follow-up. Nothing owed.
• model_builders: refuse an input that is set but missing; record capped-fit timing #832 — model_builders refuses an input that is set but missing. Nothing owed.
• stage_hg002: stop a SIGPIPE in the contig check ending the job #833 — stage_hg002 SIGPIPE fix. Nothing owed.
• archive_run: find the job log through scontrol, warn when it is missing #835 — archive_run finds the job log through scontrol.
  • Verified on Delta: campaign 22801847–55 archived each job ([archive] … files: …).
• realism panel: say what the simulated genome carries #836 — the realism panel says what the simulated genome carries. Nothing owed.
• indel_context: measure indel_probability in the same pass (#746) #838 / indel_probability: record it as measured; report both fits (#746) #839 (Measure indel_probability: the numerator is collected, the denominator never was #746) — indel_probability measured on HG002 at 0.0092–0.0104; default 0.01 kept.
  • Owed (accepted): one library at the realism loci; the LOW_FRAC threshold at high depth is untested.
• Comparison page: current het/hom default, and date the stale comparisons (#840) #841, Re-freeze fidelity baselines and refresh ACCESS from the v4.0.0 campaign (#840) #843 (Re-run germline baselines measured before the v3.4/v3.5 recalibration #840) — v4.0.0 baselines re-frozen; ACCESS and comparison page refreshed.
  • Owed: the charts were edited, not rendered in a browser. Single-job numbers, except the adapter arms.
• Docs: eidolon is not on Bioconda yet; point installs at release binaries #849 — docs no longer promise a Bioconda install. Nothing owed.
• CITATION.cff: cite both NEAT papers; version 4.0.0 #850 — CITATION.cff cites both NEAT papers; version current.
  • Owed: confirm GitHub renders "Cite this repository" once this reaches main.
• Re-freeze somatic_indel_recall at 0.742: the misses are the caller's (#847) #851 (somatic_indel_recall: replicate before deciding regression vs. noise (0.696 = 16/23, one draw) #847) — somatic_indel_recall re-frozen at 0.742, with a read-level check showing eidolon plants the indels.
  • Owed (accepted): two open sites (+1 at 3 reads; a +445 bp insertion that short reads cannot show).

Release procedure

• cargo test --workspace: 1,117 passed, 0 failed (release/v4.0.1 at 9c32f03d)
• clippy --workspace --all-targets -D warnings: clean (9c32f03d)
• release-gates green on the release PR Release v4.0.1 #855: release-blockers pass, realigned-gates pass (run on 13274d70)
• scripts/release_shakeout.sh, mode SKIP_LARGE_DATA=1 (in-repo suite only): PASS (9c32f03d)
• junk-file check: no empty tracked files, no odd names
• version bumped in Cargo.toml (+ Cargo.lock), conda-recipe/meta.yaml and CITATION.cff: Release prep: v4.0.1 #853
• five binary assets on the release page (aarch64-rhel8, apple-darwin, windows-msvc.exe, linux-gnu, linux-gnu-rhel8). The downloaded linux-gnu binary ran under its download name: 4.0.1+920a7e7, an H1N1 run in 92 ms, and a missing config exits 1 with the subcommand named.
• docs site serves: docs.yml on 920a7e7e succeeded; index 200, getting-started/installation.html 200, and it shows Docs: eidolon is not on Bioconda yet; point installs at release binaries #849's new text, so the copy served is fresh
• conda sha256 PR into develop: conda sha256 for v4.0.1 #856 (a46aa992…, identical across two downloads)
• Bioconda: bioconda-recipes#67420 updated to 4.0.1 (8e5d9a85; version and sha256 only). Still awaiting review, so not published.

Not verified for this release

• Config panics: subcommands outside the three config loaders were not audited case by case (Bad configs return errors instead of panicking #842).
• Delta timings: E. coli and yeast after the fix, on a confirmed-healthy filesystem. Only chr22 is quoted (ACCESS §3.2, ACCESS §3.2: the startup cost was the reader, fixed in v4.0.1 (#846) #854). The remaining per-base slowdown against v1.19.1 (11–16% on the same node) is Per-base generation is 11-16% slower than v1.19.1 (same node, local disk); thread scaling is not regressed #848, not a v4.0.1 change.
• Memory: the ~40 MB of fixed RSS (Peak RSS carries a fixed ~40 MB overhead that postdates the perf baseline #670).
• CITATION.cff rendering on GitHub's "Cite this repository" box (CITATION.cff: cite both NEAT papers; version 4.0.0 #850), now on main, has not been looked at.
• Release binaries: only the linux-gnu one was run. The other four were counted, not executed.
• Bioconda: #67420 is updated to 4.0.1 but still unreviewed, so eidolon is not installable via conda.